Example: How to Write Parallel Queries in Parallel 1 ECONDO

Transcription

Example: How to Write Parallel Queries in Parallel 1 ECONDO
Example: How to Write Parallel Queries in Parallel S ECONDO
Jiamin Lu, Oct 2012
S ECONDO was built by the S ECONDO team
1 Overview of Parallel Secondo
Parallel S ECONDO extends the scalability of S ECONDO database system, in order to process moving objects data
in parallel on a computer cluster. It is composed by Hadoop and a set of single-computer S ECONDO databases.
Here Hadoop is used as a task coordinator and communication level, while most database queries are processed by
S ECONDO databases simultaneously. The functions of Parallel S ECONDO are provided by two algebras, Hadoop
and HadoopParallel. Both have been published since S ECONDO 3.3.2.
The basic infrastructure of Parallel S ECONDO is shown in Figure 1. HDFS is a distributed file system prepared
by Hadoop to shuffle its intermediate data during parallel procedures. PSFS (Parallel S ECONDO File System)
is a similar system provided in Parallel S ECONDO, directly exchanging data among single-computer S ECONDO
databases without passing through HDFS, in order to improve the data transfer efficiency in the parallel system.
The minimum execution unit in Parallel S ECONDO is called Data Server (DS). Each DS contains a compact
S ECONDO named Mini-S ECONDO, its database storage and a PSFS node. At present, it is quite common that
a low-end PC has several hard disks, large memory and several processors with multiple cores. Therefore, it
is possible for the user to set several DSs on one computer, in order to fully use the computing resource of the
underlying cluster. Hadoop nodes are set independently with DSs, each is set inside the first DS of a computer.
More details about Parallel S ECONDO can be found in the document “User Guide For Parallel S ECONDO”, which
is also openly published on our website.
At the current stage, all parallel queries in Parallel S ECONDO are expressed in S ECONDO executable language. In
this document, the way of changing a sequential query into a parallel query is demonstrated, in order to help the
user quickly getting familiar with this system.
2 Install and Uninstall Parallel Secondo
Parallel S ECONDO can be installed on either a single computer or a cluster composed by tens and even hundreds
of computers. In this demonstration, Parallel S ECONDO is installed on a computer having one AMD Phenom(tm)
II X6 1055T processor with six cores, 8 GB memory and one 500 GB hard disk. Here the Ubuntu 10.04.2-LTSDesktop-64bit is used as the operating system, while other platforms including Ubuntu 12.04 and MacOSX 10.5
have also been tested.
At the same time, a virtual machine (VM) image of a single-computer Parallel S ECONDO has also been provided
on our website. The user can experience Parallel S ECONDO with the VM image without installing it on his own
machine.
1
Master Data Server
Master
Node
Mini Secondo
Meta
Database
DS Catalog
Parallel
Query
Converter
Hadoop Master Node
H
D
F
S
Slave
Node
Hadoop Slave Node
Partial Secondo
Query
DS Catalog
Slave Data Server
P
S
F
S
Mini Secondo
Slave
Database
Slave Data
Server
Slave Data
Server
...
Slave Node
... ...
Figure 1: The Infrastructure of Parallel Secondo
Prerequisites
Few utilities must be prepared before installing Parallel S ECONDO. They are all ordinary linux commands working
on many Linux platforms and MacOSX.
• JavaTM 1.6.x or above. The openjdk-6 is automatically prepared along with the installation of S ECONDO.
• SSH connection. Both Data Servers and the underlying Hadoop platform rely on secure shell as the basic
communication level. For example on Ubuntu, the SSH server is not installed by default, and can be
installed with the command:
$ sudo apt-get install ssh
• screen is also requested by Parallel S ECONDO scripts. In Ubuntu, it can be installed like:
$ sudo apt-get install screen
• Particularly in Hadoop, a passphraseless SSH connection is required. The user can check and set it up
through the following commands. First, carry out the “ssh” command to see wether a passphrase is required.
$ ssh <IP>
Here the “IP” is the IP address of the current computer, and usually can be found out through the “ifconfig”
command. If a password is asked for this connection, then an authentication key-pair should be created with
the commands:
$ cd $HOME
$ ssh-keygen -t dsa -P '' -f ˜/.ssh/id_dsa
$ cat ˜/.ssh/id_dsa.pub >> ˜/.ssh/authorized_keys
Afterward, try to ssh to the local computer again, this time it may asks to add its current IP address to the
known hosts list, like:
2
$ ssh <IP>
The authenticity of host '... (...)' can't be established.
RSA key fingerprint is .......
Are you sure you want to continue connecting (yes/no)?
This step happens only once when the ssh connection is built at the first time. The user can simply confirm
the authentication by typing “yes”. If the user prefers to avoid this step, the following three lines can be
added into the file $HOME/.ssh/config.
Host *
StrictHostKeyChecking
UserKnownHostsfile
2.1
no
/dev/null
Installation Steps
A set of auxiliary tools, which practically are bash scripts, are provided to help the user to install and use Parallel
S ECONDO easily. These tools are kept in the Hadoop algebra of S ECONDO 3.3.2, in a folder named clusterManagement. The installation of Parallel S ECONDO on a single computer includes the following steps:
1. Install S ECONDO. The installation guide of S ECONDO for different platforms can be found on our website,
and the user can install it as usual.1 Afterward, the user can verify the correctness of the installation and
compile S ECONDO.
$ env | grep 'ˆSECONDO'
SECONDO_CONFIG=.... /secondo/bin/SecondoConfig.ini
SECONDO_BUILD_DIR=... /secondo
SECONDO_JAVA=.... /java
SECONDO_PLATFORM=...
$ cd $SECONDO_BUILD_DIR
$ make
Particularly, in Ubuntu, the following line in the profile file $HOME/.bashrc
source $HOME/.secondorc $HOME/secondo
should not be set at the end of the file. Instead, it should be set above the line:
[ -z "$PS1" ] && return
2. Download Hadoop. Go to the official website of Hadoop, and download the Hadoop distribution with the
version of 0.20.2. The downloaded package should be put into the $SECONDO BUILD DIR/bin directory
without changing the name.
$ cd $SECONDO_BUILD_DIR/bin
$ wget http://archive.apache.org/dist/hadoop/core/
hadoop-0.20.2/hadoop-0.20.2.tar.gz
3. A profile file named ParallelSecondoConfig.ini is prepared for setting all parameters in Parallel S ECONDO.
Its example file is kept in the clusterManagement folder of the Hadoop Algebra, which is basically made
for a single computer with Ubuntu. However, few parameters still need to be set, according to the user’s
computer.
1
The version must be 3.3.2 or higher.
3
• The Cluster
like:
2
parameter indicates the DSs in Parallel S ECONDO. In a single computer, it can be set
Master = <IP>:<DS_Path>:<Port>
Slaves += <IP>:<DS_Path>:<Port>
The “IP” is the IP address of the current computer. The “DS Path” indicates the partition for a Data
Server, for example /tmp. Note that the user must have the read and write access to the DS Path. The
“Port” is also a port number prepared for a DS daemon, like 11234. Different DS can be set on the
same computer, but their DS Paths and Ports must be different.
• Set hadoop-env.sh:JAVA HOME to the location where the JAVA SDK is installed,
• The user might already have installed S ECONDO before, and created some private data in the database.
If so, the NS4Master parameter can be set true, in order to let Parallel S ECONDO visit the existing
databases.
NS4Master = true
Note here if there is no S ECONDO databases created before, and the NS4Master parameter is set to be
true, then the below installation will fail.
• At last, the transaction feature is normally turned off in Parallel S ECONDO, in order to improve the
efficiency of exchanging data among DSs. For this purpose, the following RTFlag parameter in the
S ECONDO configuration file should be uncommented.
RTFlags += SMI:NoTransactions
The file is named SecondoConfig.ini, being kept in the $SECONDO BUILD DIR/bin.
After setting all required parameters, copy the file ParallelSecondoConfig.ini to $SECONDO BUILD DIR/bin,
and start the installation with the auxiliary tool ps-cluster-format.
$ cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
$ cp ParallelSecondoConfig.ini $SECONDO_BUILD_DIR/bin
$ ps-cluster-format
During this installation, all Data Servers will be created and the Hadoop is installed. Besides, the Namenode
in Hadoop is formatted at last.
4. Close the shell and start a new one. Verify the correctness of the initialization with the following command:
$ cd $HOME
$ env | grep 'ˆPARALLEL_SECONDO'
PARALLEL_SECONDO_MASTER=.../conf/master
PARALLEL_SECONDO_CONF=.../conf
PARALLEL_SECONDO_BUILD_DIR=.../secondo
PARALLEL_SECONDO_MINIDB_NAME=msec-databases
PARALLEL_SECONDO_MINI_NAME=msec
PARALLEL_SECONDO_PSFSNAME=PSFS
PARALLEL_SECONDO_DBCONFIG=.../SecondoConfig.ini....
PARALLEL_SECONDO_SLAVES=.../conf/slaves
PARALLEL_SECONDO_MAINDS=.../dataServer1/...
PARALLEL_SECONDO=.../dataServer1
PARALLEL_SECONDO_DATASERVER_NAME=...
2
This parameter must be set before continue.
4
5. The third step initializes the DS in the computer, but the Mini-S ECONDO system is not distributed into those
Data Servers yet. This is because the Hadoop algebra cannot be compiled before setting up the environment
of Parallel S ECONDO. Now both Hadoop and HadoopParallel algebras should be activated, and S ECONDO
should be recompiled. The new algebras are activated by adding the following lines to the algebra list file
$SECONDO BUILD DIR/makefile.algebras.
ALGEBRA_DIRS += Hadoop
ALGEBRAS
+= HadoopAlgebra
ALGEBRA_DIRS += HadoopParallel
ALGEBRAS
+= HadoopParallelAlgebra
After re-compiling the S ECONDO system, the user can distribute Mini-S ECONDO to all local DSs with the
auxiliary tool ps-secondo-buildMini.
$
$
$
$
cd $SECONDO_BUILD_DIR
make
cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
ps-secondo-buildMini -lo
So far, Parallel S ECONDO has been installed. If the user wants to completely remove it from the computer or the
cluster, an easy-to-use tool ps-cluster-uninstall is also provided.
$ cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
$ ps-cluster-uninstall
Besides installing Parallel S ECONDO on the user’s own computer, a VMware image containing a single-computer
Parallel S ECONDO is also provided, with which the user can easily prepare the system by starting a virtual machine.
3 Start and Shutdown Parallel Secondo
In Parallel S ECONDO, Hadoop and DSs are deployed independently, therefore they also are started and closed
with separate steps. Hadoop is started by its own script start-all.sh, while the DSs are started by ps-startMonitors.
This script starts all monitors, i.e. daemons of Mini-S ECONDO databases distributed in DSs. Besides, the user can
apply the ps-cluster-queryMonitorStatus to check the status of all involved Mini-S ECONDO.
$
$
$
$
$
start-all.sh
cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
ps-startMonitors
ps-cluster-queryMonitorStatus
ps-startTTYCS -s 1
At last, the user opens a text interface of Parallel S ECONDO with the script ps-startTTYCS. This interface connects
to the Mini-S ECONDO database in the master DS, which is the first DS in the master node of the cluster. Here the
Parallel S ECONDO is installed on a single computer with one DS only, thus the s argument is set as one.
Besides opening the text interface, the user can also start up the graphical interface provided by S ECONDO, and
connect with the master DS by setting the IP address and the access port that we set in the ParallelSecondoConfig.ini file.
Parallel S ECONDO can be shut down by separately stopping all Mini-S ECONDO monitors and the Hadoop framework.
5
$ stop-all.sh
$ cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
$ ps-stopMonitors
4 A Simple Parallel Query Example
In order to make the user familiar with Parallel S ECONDO as soon as possible, a simple example is offered first.
Start the text interface of Parallel S ECONDO, and then load an example database named opt:
restore database opt from opt;
In this example database, there prepared two simple relations: plz and Orte. The plz has two attributes: Ort (string)
and PLZ (int), recording zip codes of some German cities. Orte describes some cities’ further information, except
the post code. Both relations use the string attribute Ort for city names. The following query performs a hash join
operation on these two relations:
query plz feed {p} Orte feed {o} hashjoin[Ort_p, Ort_o] count;
10052
In the below of this section, the parallel data model in Parallel S ECONDO and the method of changing a sequential
query to a parallel query are introduced.
4.1
Parallel Data Model
In order to describe the parallel procedure of the above example, the Parallel data model is provided. It helps
the user to write parallel queries like ordinary sequential queries. The data type flist encapsulates all S ECONDO
objects and distributes their content on the computer cluster. Besides, a set of parallel operators are implemented
in Parallel S ECONDO, making the system compatible with the single-computer database. These operators can be
roughly divided into four kinds: flow operators, Hadoop operators, PSFS operators, and other assistant operators,
as shown in Table 1.
Flow operators connect parallel operators with other existing S ECONDO operators. Relations can be distributed
over the cluster with flow operations, and then be processed by all DSs in parallel. Later the distributed result,
which is represented as an flist object, can be collected from the cluster in order to be processed with singlecomputer S ECONDO. Hadoop operators accept queries written in S ECONDO executable language and create
Hadoop jobs with their respective template Hadoop programs. Thereby the user can write parallel queries like
common sequential queries. Assistant operators work together with Hadoop operators, helping to denote flist
objects in parameter functions. At last, PSFS operators are particularly prepared for accessing data exchanged in
PSFS, being internally used in Hadoop operations. Details about these operators can be found in the User Guide
of Parallel S ECONDO.
4.2
The Parallel Solution
With the Parallel data model, the example join query can be processed in parallel with following queries.
let CLUSTER_SIZE = 1;
let PS_SCALE = 6;
let plz_p = plz feed spread[; Ort, CLUSTER_SIZE, TRUE;];
6
Type
Flow
Hadoop
Assistant
PSFS
Operator
spread
collect
hadoopMap
hadoopReduce
hadoopReduce2
para
fconsume
fdistribute
ffeed
Signature
stream(T) →flist(T)
flist(T) →stream(T)
flist x fun(mapQuery) x bool →flist
flist x fun(reduceQuery) →flist
flist x flist x fun(reduceQuery) →flist
flist(T) →T
stream(T) x fileMeta →bool
stream(T) x fileMeta x Key→stream(key,tupNum)
fileMeta →stream(T)
Table 1: Operators Provided as Extensions In Parallel Secondo
let Orte_p = Orte feed spread[; Ort, CLUSTER_SIZE, TRUE;];
query plz_p Orte_p hadoopReduce2[ Ort, Ort, DLF, PS_SCALE
; . {p} .. {o} hashjoin[Ort_p, Ort_o] ]
collect[] count;
10052
At the beginning the queries, two constants CLUSTER SIZE and PS SCALE are defined. The CLUSTER SIZE
tells the number of slave DS, hence here it is set as one. The constant PS SCALE indicates the number of tasks
running in parallel during the Hadoop procedure. It is set to six as there are six processor cores in the current
computer.
Next, both flist objects plz p and Orte p are created with the spread operator. The spread distributes a relation
into pieces on the cluster, based on an indicated attribute. Here the first argument in both queries is Ort, hence
both relations are divided by the Ort attribute. The second argument indicates the scale of the partition. Usually it
is set according to the number of the slave DSs, i.e. the CLUSTER SIZE. The third argument is true, in order to
keep the partition attribute in distributed data pieces.
At last, the hadoop operator hadoopReduce2 is used to process the parallel query. Both relations are repartitioned based on the Ort attribute, with the scale up to PS SCALE. The result is set as a DLF flist, in order to
be collected into the master DS with the collect operator. The argument function set in the hadoopReduce2
defines the database query being executed in every reduce task. As you can see, it is the same as the sequential
query.
The result of the parallel query is correct, but it takes a much longer time than the sequential query. This is normal
since there exist constant overhead for carrying out a Hadoop job. Usually parallel procedures are prepared for i/oand cpu- intensive queries dealing with a large-scale data set, like the one we will introduce in the next section.
5 A Spatial Query Example
In this section, a more complex spatial query is introduced and adapted for parallel processing. It studies the
problem of dividing a massive amount of large regions into small pieces, based on a set of polylines that overlap
them. For example, a forest is split into pieces by roads. Regarding this issue, an operator findCycles is provided
in S ECONDO, and used in the following query:
query Roads feed {r}
Natural feed filter[.Type contains 'forest']
itSpatialJoin[GeoData_r, GeoData, 4, 8]
7
sortby[Osm_id]
groupby[Osm_id; InterGraph:
intersection( group feed projecttransformstream[GeoData_r]
collect_line[TRUE], group feed extract[GeoData])
union boundary(group feed extract[GeoData])]
projectextendstream[Osm_id; Curves: findCycles(.InterGraph)]
count;
Total runtime ...
Times (elapsed / cpu): 22:53min (1372.55sec)
/899.94sec = 1.52516
195327
Here two relations Natural are Roads are used. Both are public data sets fetched from the OpenStreetMap project,
describing the geographic information of the German state North Rhine Westphalia (NRW). They can be downloaded from the website and then imported into the S ECONDO database as relations with queries that will shown
at the end of this section.
The Natural relation contains many region shapes. They are classified into four kinds according to their actual
usages, including forest, park, water and riverbank, indicated by the attribute Type. Shapes in the Roads relation
are polylines, consisting of the street network in that area. Both relations use the geometric attribute GeoData
to record their precise shape descriptions, and each shape is identified by the key attribute Osm id. When this
document is written, the Natural relation contains 92,875 regions, from which 71,415 (77%) are forests. Its size
is about 421 MB. At the same time, there are in total 1,153,587 roads listed in Roads, which is as large as 1.5 GB.
Note that the data sets on OpenStreetMap are updated aperiodically, therefore the scale of these data sets and also
the above query result may change along with the upgrade of the data sets.
In the example query, the forest regions in the relation Natural are split into pieces by the road network from
the relation Roads. The query is written in S ECONDO executable language, consisting of S ECONDO objects and
operators. It describes the data flow of the query, and the user can use different methods and operators for the
same query.
This example query is mainly divided into three steps. At first, forest regions in the relation Natural are joined
with the relation Roads, by using the itSpatialJoin operator. The itSpatialJoin is the most efficient sequential
spatial join operation in the current S ECONDO. It builds up a memory R-tree on one side relation, and probes this
R-tree with the tuples from the other relation. Pair-tuples which have overlapped bounding boxes are output as
the join result, which then be grouped based on forest regions’ identifiers. For each group, a graph is built up
consisting of all boundary edges of the region and all its intersecting roads. For example, as shown in Figure 2,
the region R is divided into five pieces by three intersected lines a, b and c, and the created graph is shown in
Figure 2b.
First, all roads intersecting the region are put together into one line value by the collect line operator. The
intersection operator then computes all parts of this line (road lines) lying within the forest region, like P1P2,
P3P4, P7P8, etc. The boundary operator produces all boundary edges of the region, like P5P6. Boundary edges
and roads inside the region are put together by the union operation.
At last, the findCycles operator traverses each graph (represented as a line value) and returns split regions, like
the red cycle shown in the Figure 2b. It is used as an argument function in the operator projectextendstream,
thereby each forest region’s identifier is duplicated for all its split regions. This sequential query produces 195,327
split regions and costs about 1373 seconds.
5.1
Import Spatial Relations
Both the Natural and Roads relations are imported from the OpenStreetMap project with the following steps.
8
(a)
(b)
Figure 2: findCycles Operation
$ cd $SECONDO_BUILD_DIR/bin
$ wget http://download.geofabrik.de/europe/germany/
nordrhein-westfalen-latest.shp.zip
$ mkdir nrw
$ unzip -d nrw nordrhein-westfalen.shp.zip
By now, all shape files about streets and regions in NRW are kept in the directory named nrw in the $SECONDO BUILD DIR/bin. The user can then use the following S ECONDO queries to load them into the database,
and the elapsed time of partial queries are listed below.
create database nrw;
open database nrw;
let Roads =
dbimport2('../bin/nrw/roads.dbf') addcounter[No, 1]
shpimport2('../bin/nrw/roads.shp') namedtransformstream[GeoData]
addcounter[No2, 1]
mergejoin[No, No2] remove[No, No2]
filter[isdefined(bbox(.GeoData))] validateAttr
consume;
Total runtime ...
Times (elapsed / cpu): 2:08min (128.28sec)
/90.49sec = 1.41761
let Natural =
dbimport2('../bin/nrw/natural.dbf') addcounter[No, 1]
shpimport2('../bin/nrw/natural.shp') namedtransformstream[GeoData]
addcounter[No2, 1]
mergejoin[No, No2] remove[No, No2]
filter[isdefined(bbox(.GeoData))] validateAttr
consume;
Total runtime ...
Times (elapsed / cpu): 39.1086sec / 30.38sec = 1.28731
close database;
9
Here we use the relative paths to indicate the data and shape files that the dbimport2 and shpimport2 operators need. In case they cannot find these files for various reasons, the user can replace these relative paths with
their absolute paths.
5.2
The Parallel Solution
The above sequential spatial query can be processed in parallel by adopting the idea of semi-join, consisting of
the queries shown in the following:
open database nrw;
let CLUSTER_SIZE = 1;
let PS_SCALE = 6;
let JoinedRoads_List = Roads feed
Natural feed filter[.Type contains 'forest'] {r}
itSpatialJoin[GeoData, GeoData_r, 4, 8]
projectextend[Osm_id, GeoData; OverRegion: .Osm_id_r]
spread[; OverRegion, CLUSTER_SIZE, TRUE;]
Total runtime ...
Times (elapsed / cpu): 9:03min (542.65sec)
/82.81sec = 6.55295
let Forest_List = Natural feed filter[.Type contains 'forest']
spread[; Osm_id, CLUSTER_SIZE, TRUE;]
Total runtime ...
Times (elapsed / cpu): 10.8835sec
/ 3.62sec = 3.00649
query JoinedRoads_List Forest_List
hadoopReduce2[OverRegion, Osm_id, DLF, PS_SCALE
; . {r} .. hashjoin[OverRegion_r, Osm_id]
sortby[Osm_id] groupby[Osm_id
; InterGraph: intersection( group feed projecttransformstream[GeoData_r]
collect_line[TRUE], group feed extract[GeoData])
union boundary(group feed extract[GeoData])]
projectextendstream[Osm_id; Curves: findCycles(.InterGraph)]]
collect[] count;
Total runtime ...
Times (elapsed / cpu): 7:01min (420.704sec)
/4.31sec = 97.6111
195327
close database;
At first, the constants CLUSTER SIZE and PS SCALE should be set in the nrw database. The first sequential
query generates all roads that intersect with forest regions, with the itSpatialJoin operator. In the mean time, if
a road passes several forests, it is duplicated for every intersection. At last, each road tuple is only extended with
the identifiers of the forest regions that it overlaps, by using the projectextend operator.
The parallel query is mainly built up with the hadoopReduce2 operator. It first distributes forest regions and
all their joined roads based on forest identifiers, hence in each reduce task, forest-road pairs can be created with
a hashjoin operation. Thereafter, forest regions are grouped with all roads that it intersects by the groupby
10
operator, and the following query is just the same as the sequential solution. In the end, the distributed results are
collected back to the standard S ECONDO by using the flow operator collect.
This complete solution costs 975 seconds in total, and gains about 1.4 times speed-up in a single computer.
A
Alternate Parallel Approaches
Till now, queries in Parallel S ECONDO are represented in S ECONDO executable query language, with which the
user can choose different combinations of operators to process the same query.
At the same time, in total there are three Hadoop operators provided in Parallel S ECONDO, as shown in the
Table 1, describing parallel queries processed in the Map and Reduce stages of Hadoop jobs, respectively. The
hadoopReduce2 operator demonstrated above processes the query, which is described as an argument function,
in the Reduce stage. The hadoopReduce accepts only one input flist object, but also process the query in the
Reduce stage. On the contrary, the hadoopMap processes its argument function by Map tasks only. With these
Hadoop operators, there are two alternate parallel methods also can process the example spatial query that we
presented above. Although they are not as efficient as the approach that we introduced, it indicates that Parallel
S ECONDO can flexibly process the user’s various custom problems with different parallel solutions.
Since the two alternate solutions are not efficient enough being processed on a single computer, we would like to
introduce and evaluate them in a computer cluster. This cluster contains six computers like the one that we used
before, but each computer has two hard drives, thus we can set up two Data Servers on every computer. Hereby,
the CLUSTER SIZE is set as twelve, and the PS SCALE is thirty six since we have so many cores in the whole
cluster. In the mean time, we evaluate the former parallel solution on the cluster as well.
A.1
Install Parallel Secondo on a Cluster
Installing Parallel S ECONDO on a cluster is almost the same as the installation on a single computer. The auxiliary
tools help the user to set up and use the system easily, with only the configure file need to be adjusted.
Prerequisites
Before the installation, there are some basic environment that the cluster should prepare. First, all computers in
the cluster should have a same operating system, or at least all of them should be Linux or Mac OSX. Second, a
same account needs to be created on all machines. This can be easily done with services like NIS. Thereafter, the
SSH and screen services should be prepared on all computers, and the computers can visit each other through SSH
without using the passphrase. Third, one or several disk partitions should be prepared on every machine, and the
account prepared before should have read and write access on them. At last, all required libraries of S ECONDO
and the .secondorc file should be installed on every computer. Likewise, if the operating system is Ubuntu, then
the line
source $HOME/.secondorc $HOME/secondo
should be set at the top of the profile file $HOME/.bashrc, before
[ -z "$PS1" ] && return
Installation Steps
One computer of the cluster must be indicated as the master node, and the complete installation is done on this
machine only, with the following steps:
11
1. Enter the master node, download the latest S ECONDO distribution, which must be newer than 3.3.2. Then
extract the S ECONDO archive to the master’s $SECONDO BUILD DIR, and compile it. Afterward, download the required Hadoop archive to $SECONDO BUILD DIR/bin.
$
$
$
$
$
tar -xzf secondo-v33*.tar.gz
cd $SECONDO_BUILD_DIR
make
cd $SECONDO_BUILD_DIR/bin
wget http://archive.apache.org/dist/hadoop/core/
hadoop-0.20.2/hadoop-0.20.2.tar.gz
2. Prepare the ParallelSecondoConfig.ini file, according to the environment of the cluster. For example, the
cluster can be set as:
Master
Slaves
Slaves
Slaves
Slaves
= 192.168.1.1:/disk1/dataServer1:11234
+= 192.168.1.1:/disk1/dataServer1:11234
+= 192.168.1.2:/disk1/dataServer1:11234
+= 192.168.1.1:/disk2/dataServer2:14321
+= 192.168.1.2:/disk2/dataServer2:14321
Here a cluster with two computers and four DSs are described. The two computers are set with the IP
address of 192.168.1.1 and 192.168.1.2, respectively. Since every computer has two disks, each is set with
two DSs. The computer 192.168.1.1 is set to be the master node of the cluster, and its first DS with the port
of 11234 is set to be the master and the slave DS at the same time. Besides, the NS4Master can also be set
as true, if the user want to visit the existing S ECONDO database on the master node.
Particularly, the following five parameters in the parallel configure file should be changed, by replacing the
localhost with the master node’s IP address.
core-site.xml:fs.default.name = hdfs://192.168.1.1:49000
hdfs-site.xml:dfs.http.address = 192.168.1.1:50070
hdfs-site.xml:dfs.secondary.http.address = 192.168.1.1:50090
mapred-site.xml:mapred.job.tracker = 192.168.1.1:49001
mapred-site.xml:mapred.job.tracker.http.address = 192.168.1.1:50030
Additionally, the transaction feature should also be disabled by uncommenting the line in the file $SECONDO BUILD DIR/bin/SecondoConfig.ini on the master node.
RTFlags += SMI:NoTransactions
At last, run the ps-cluster-format script.
$ cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
$ cp ParallelSecondoConfig.ini $SECONDO_BUILD_DIR/bin
$ ps-cluster-format
3. Activate the Hadoop and HadoopParallel algebras into the file makefile.algebras, recompile S ECONDO, and
distribute Mini-S ECONDO to all DSs at last.
$
$
$
$
cd $SECONDO_BUILD_DIR
make
cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
ps-secondo-buildMini -co
The parameter c in the ps-secondo-buildMini indicates the whole cluster.
12
Till now, Parallel S ECONDO is completely installed. Accordingly, the user can also easily remove the system
with the ps-cluster-uninstall script. The user can start all S ECONDO monitors with the script ps-start-AllMonitors,
and close them with ps-stop-AllMonitors. The text interface of the whole Parallel S ECONDO is still opened by
connecting to the master DS. All these steps should, and only need to be done on the master node.
$
$
$
$
$
start-all.sh
cd $SECONDO_BUILD_DIR/Algebras/Hadoop/clusterManagement
ps-start-AllMonitors
ps-cluster-queryMonitorStatus
ps-startTTYCS -s 1
Sometimes, it is quite difficult for the user to build up a physical cluster. Regarding this issue, a Amazon AMI of
Parallel S ECONDO is provided. With this image, the user can quickly set up a virtual cluster on the Amazon Web
Services (AWS), and the Parallel S ECONDO is already built inside it, hence he/she can test the system directly.
A.2
The Former Method
The former method introduced in Section 5.2 can of course be evaluated again in the cluster. However, on a single
computer, its first query is done sequentially, taking a considerable overhead. Since here the method is carried
out on the cluster, this spatial join should also be processed in a parallel way. The parallel spatial join adopts
the method named PBSM (Partition-Based Spatial Merge). It first partitions both roads and forest regions into
a coarsely divided cell-grid, and then processes the join operation independently within each cell. At the same
time, the operator gridintersects removes duplicate join results produced in different cells. The generation of
the cell-grid is demonstrated in the Section B, and must be carried out before the following queries.
open database nrw;
let Roads_BCell_List = Roads feed extend[Box: bbox(.GeoData)]
projectextendstream[Osm_id, GeoData, Box
; Cell: cellnumber(.Box, Big_RNCellGrid)]
spread[; Cell, CLUSTER_SIZE, TRUE;];
Total runtime ...
Times (elapsed / cpu): 1:24min (83.5428sec)
/32.63sec = 2.56031
let Forest_BCell_List = Natural feed filter[.Type contains 'forest']
extend[Box: bbox(.GeoData)]
projectextendstream[Osm_id, GeoData, Box
; Cell: cellnumber(.Box, Big_RNCellGrid)]
spread[;Cell, CLUSTER_SIZE, TRUE;];
Total runtime ...
Times (elapsed / cpu): 18.1649sec
/ 4.03sec = 4.50742
let JoinedRoads_List = Roads_BCell_List Forest_BCell_List
hadoopReduce2[Cell, Cell, DLF, PS_SCALE
; . .. {r} itSpatialJoin[GeoData, GeoData_r, 4, 8]
filter[ (.Cell = .Cell_r) and
gridintersects(Big_RNCellGrid, .Box, .Box_r, .Cell)]
projectextend[Osm_id, GeoData; OverRegion: .Osm_id_r] ];
Total runtime ...
/0.1sec = 737.165
Times (elapsed / cpu): 1:14min (73.7165sec)
13
let Forest_List = Natural feed filter[.Type contains 'forest']
spread[; Osm_id, CLUSTER_SIZE, TRUE;]
Total runtime ...
Times (elapsed / cpu): 20.6345sec
/ 4.47sec = 4.61623
query JoinedRoads_List Forest_List
hadoopReduce2[OverRegion, Osm_id, DLF, PS_SCALE
; . {r} .. hashjoin[OverRegion_r, Osm_id]
sortby[Osm_id] groupby[Osm_id; InterGraph:
intersection( group feed projecttransformstream[GeoData_r]
collect_line[TRUE], group feed extract[GeoData])
union boundary(group feed extract[GeoData])]
projectextendstream[Osm_id; Curves: findCycles(.InterGraph)]]
collect[] count;
Total runtime ...
/13sec = 8.0899
Times (elapsed / cpu): 1:45min (105.169sec)
195327
The top three queries achieve the parallel spatial join with PBSM, and generate all roads that intersect with forest
regions. The other two queries are the same as we used before. This time, the whole processing costs 302 seconds
in total, earning a speed-up of nearly factor five. It turns out that this solution fits the example spatial problem
well, both on a single computer and on a cluster.
A.3
Method 1
This solution divides the example query into two stages, Map and Reduce, following the programming paradigm
of Hadoop. The spatial join operation is first finished in the Map stage, again with the PBSM method, then the
join results are distributed over the cluster based on forest identifiers, and the remaining steps are processed on all
slaves simultaneously in the Reduce stage. The queries are listed below.
let Roads_BCell_List = Roads feed extend[Box: bbox(.GeoData)]
projectextendstream[Osm_id, GeoData, Box
; Cell: cellnumber(.Box, Big_RNCellGrid)]
spread[; Cell, CLUSTER_SIZE, TRUE;];
Total runtime ...
Times (elapsed / cpu): 1:24min (83.5428sec)
/32.63sec = 2.56031
let Forest_BCell_List = Natural feed filter[.Type contains 'forest']
extend[Box: bbox(.GeoData)]
projectextendstream[Osm_id, GeoData, Box
; Cell: cellnumber(.Box, Big_RNCellGrid)]
spread[;Cell, CLUSTER_SIZE, TRUE;];
Total runtime ...
Times (elapsed / cpu): 18.1649sec
/ 4.03sec = 4.50742
query Roads_BCell_List
hadoopMap[DLF, FALSE
; . {r} para(Forest_BCell_List)
itSpatialJoin[GeoData_r, GeoData, 4, 8]
14
filter[ (.Cell = .Cell_r) and
gridintersects(Big_RNCellGrid, .Box_r, .Box, .Cell_r)]]
hadoopReduce[Osm_id, DLF, PS_SCALE
; . sortby[Osm_id]
groupby[Osm_id; InterGraph:
intersection( group feed projecttransformstream[GeoData_r]
collect_line[TRUE], group feed extract[GeoData])
union boundary(group feed extract[GeoData])]
projectextendstream[Osm_id; Curves: findCycles(.InterGraph)]]
collect[] count;
Total runtime ...
Times (elapsed / cpu): 12:56min (775.609sec)
/15.86sec = 48.9034
193527
Here the hadoopMap and hadoopReduce operators are used together in one parallel query, and each one
describes one stage of the Hadoop job. In principle, one Hadoop operator starts a Hadoop job, and processes
its argument function in one of the two stages. Nevertheless, an argument named executed is provided in the
hadoopMap operator. If this argument is set false, then the Hadoop job does not start, and the argument
function will be put off to the next Hadoop job represented with a hadoopReduce or hadoopReduce2
operator. Therefore here both operations are actually processed with one Hadoop job in the end.
The hadoopMap is a unary operation, but the itSpatialJoin operator requires two inputs. Regarding this
issue, the assistant operator para is used to identify the distributed relation Forest BCell List. The spatial join
results produced in the Map stage are then shuffled based on the forest regions’ identifiers. We can see that the
argument function embedded in the hadoopReduce query is exactly the same as in the sequential query.
This approach is not efficient for the example problem. It costs in total 878 seconds, and the speed-up factor is
about 1.6, although it is processed with six machines. This is mainly caused by the spatial join result, which are
generated and accessed mainly in the memory during the sequential procedure. However in the parallel query,
they are generated between the Map and Reduce stage, and have to be materialized and shuffled as disk files.
Since the join result is extremely large, the overhead of exporting and loading them degrades the performance.
A.4
Method 2
This solution intends to finish both the spatial join and the findCycles operation in the Map stage, in order to
avoid the shuffle procedure. The parallel spatial join is again processed with the PBSM method.
However, during the experiment, a new problem arises. When a region overlaps several cells, not all its intersected
lines overlap the same cells. For example, as shown in Figure 3, the region is partitioned into two cells C1 and
C2. According to PBSM, it is duplicated in both cells. The lines P3P7 and P4P8 are also partitioned to C1 and
C2. However, the line P2P9 is only assigned to C1. Therefore, in the cell C2, because of the absence of P2P9, the
wrong region P1-P3-P10-P8-P9 is produced. This problem is called Over-Cell Problem.
15
Figure 3: Over-Cell Problem
let Roads_BCell_List = Roads feed extend[Box: bbox(.GeoData)]
projectextendstream[Osm_id, GeoData, Box
; Cell: cellnumber(.Box, Big_RNCellGrid)]
spread[; Cell, CLUSTER_SIZE, TRUE;];
Total runtime ...
Times (elapsed / cpu): 1:24min (83.5428sec)
/32.63sec = 2.56031
let OneCellForest_BCell_OBJ_List =
Natural feed filter[.Type contains 'forest']
extend[Box: bbox(.GeoData)]
filter[cellnumber( .Box, Big_RNCellGrid) count = 1]
projectextendstream[Osm_id, GeoData, Box
; Cell: cellnumber(.Box, Big_RNCellGrid)]
spread[; Cell, CLUSTER_SIZE, TRUE;]
hadoopMap[; . consume];
Total runtime ...
/5.61sec = 13.6308
Times (elapsed / cpu): 1:16min (76.4691sec)
query Roads_BCell_List
hadoopMap[DLF ; . {r} para(OneCellForest_BCell_OBJ_List) feed
itSpatialJoin[GeoData_r, GeoData, 4, 8]
filter[gridintersects(Big_RNCellGrid, .Box_r, .Box, .Cell_r)]
sortby[Osm_id]
groupby[Osm_id; InterGraph:
intersection( group feed projecttransformstream[GeoData_r]
collect_line[TRUE], group feed extract[GeoData])
union boundary(group feed extract[GeoData])]
projectextendstream[Osm_id; Curves: findCycles(.InterGraph)]]
collect[] count;
Total runtime ...
/7.16sec = 22.8746
Times (elapsed / cpu): 2:44min (163.782sec)
169874
let MulCellForest = Natural feed filter[.Type contains 'forest']
filter[cellnumber( bbox(.GeoData), Big_RNCellGrid) count > 1]
consume;
16
Total runtime ...
Times (elapsed / cpu): 3.26407sec
/ 1.47sec = 2.22046
query Roads feed {r} MulCellForest feed
itSpatialJoin[GeoData_r, GeoData, 4, 8]
sortby[Osm_id] groupby[Osm_id; InterGraph:
intersection( group feed projecttransformstream[GeoData_r]
collect_line[TRUE], group feed extract[GeoData])
union boundary(group feed extract[GeoData])]
projectextendstream[Osm_id; Curves: findCycles(.InterGraph)]
count
Total runtime ...
Times (elapsed / cpu): 5:42min (341.884sec)
/170.5sec = 2.00519
25453
Regarding the Over-Cell problem, the forest regions are divided into two types, those covering only one cell of
the grid and those covering multiple cells. For the first type, we process them in Parallel S ECONDO, while the
second type is still processed in the sequential query.
In particular, a hadoopMap operation is used to load all forest regions that cover only one cell into distributed
S ECONDO databases, creating a DLO type flist object at last. This is because such an object is used by the
assistant operator para in the parallel query, each task can use part of its value only when it is a DLO flist.
This solution costs in total 668 seconds, achieving a speed-up factor of two. Although the result is correct, the
efficiency is not satisfying. At the same time, because of the Over-Cell problem, the solution has to be divided to
two steps, making the user not so easy to understand.
B
Assistant Objects
Here the cell-grid used in the above PBSM method is created, named Big RNCellGrid. The cell number is set
to the number of DSs, i.e., CLUSTER SIZE. This grid is coarsely divided, since the efficient itSpatialJoin
operation is executed in every cell.
At present, the two relations have to be completely scanned so as to obtain the bounding box of the grid. The grid
is created by an operator named createCellGrid2D, requiring five arguments. The first two arguments indicate
the bottom-left point of the grid, while the other three define the size of cell edges and the cell number on the
X-axis.
let CLUSTER_SIZE = 12;
let PS_SCALE = 36;
let RoadsMBR = Roads feed
extend[Box: bbox(.GeoData)]
aggregateB[Box; fun(r:rect, s:rect) r union s
; [const rect value undef]];
let NaturalMBR = Natural feed
extend[Box: bbox(.GeoData)]
aggregateB[Box; fun(r:rect, s:rect) r union s
; [const rect value undef]];
17
let RN_MBR = intersection(RoadsMBR, NaturalMBR);
let BCellNum = CLUSTER_SIZE;
let BCellSize = (maxD(RN_MBR,1) - minD(RN_MBR,1)) / BCellNum;
let Big_RNCellGrid =
createCellGrid2D(minD(RN_MBR,1), minD(RN_MBR,2),
BCellSize, BCellSize, BCellNum);
18